Logo

Harvard ML theory talk - Stefanie Jegelka - Shared screen with speaker view
Wei Lu
02:22:38
Should it be m^k_v in step 2?
vladan
02:22:49
Do we use message from (k-1)-th or k-th step in combine?
Hamza Chaudhry
02:23:25
Do these Aggregate/Combine operations take the node degree into account? Since high degree nodes would capture information from the entire graph more quickly than low degree nodes.
Mehmet Yildirim
02:24:41
Should we think each feature as a iteration?
Vicky Jiaqi Zhang
02:36:29
What’s the latent dimension?
Mark Kong
02:36:39
Did we choose to have weights shared between iterations?
Mehmet Yildirim
02:48:47
How much does the complexity of the network increase by creating a fully-connected NN?
Khalid Saifullah
03:03:02
Does the hyperbolic tangent does any better job than ReLU in terms of extrapolation?
Gautam Reddy Nallamala
03:11:28
How is the network trained end-to-end with a min function?
Bharat
03:16:11
How are we encoding the prior information that you talked about earlier about the problem we aretrying to learn so that we can do better in out of train set distribution, did we encode some priorprobability Distribution , if so how we are doing here now?
Mark Kong
03:21:06
Can you learn the linear functions directly, instead of as MLPs?
Tolga Ergen
03:21:21
Could you give more information about ReLU networks fitting linear function away from training data?
Tolga Ergen
03:23:04
No I meant the plots you showed where ReLU networks don’t extrapolate
Mark Kong
03:25:17
Is the theorem false if you don't assume that the network is trained by GD in this theorem?
Tolga Ergen
03:28:56
Thanks, can you also provide further details regarding the ReLU network structure in your experiments such as depth, width, dropout, batch norm etc.
Saptarshi Sengupta
03:30:57
Dr. Jegelka, the talk was extremely interesting! I'm quite unfamiliar with GNN's.. So, for a novice, are there any resources I could use to learn more about GNN's?
Mark Kong
03:31:48
You mentioned earlier how an RNN is like a GNN but where the update graph forms a line, but do we assume the weights in the RNN aren't necessariy shared among iterations?
Saptarshi Sengupta
03:31:48
Thank you!
Cong Peng
03:31:57
If the problem is not dynamic programing in the core, the GNN work?
Boaz Barak
03:32:16
https://www.cs.mcgill.ca/~wlh/grl_book/
Mark Kong
03:36:47
Could I have a reference to the matrix algebra thing Boaz mentioned earlier?
Boaz Barak
03:37:39
https://web.eecs.umich.edu/~pettie/papers/APBP.pdf
Boaz Barak
03:38:05
https://arxiv.org/abs/1911.06132
Cong Peng
03:38:32
Thanks!!
Dimitris Kalimeris
03:38:37
thanks!
Mehmet Yildirim
03:38:41
Thank you